A trichotomy theorem in natural models of AD +

Contemporary Mathematics A trichotomy theorem in natural models of AD + Andrés Eduardo Caicedo and Richard Ketchersid Abstract. Assume AD + and that either V = L(P(R)), or V = L(T, R) for some set T ORD. Let (X, ) be a pre-partially ordered set. Then exactly one of the following cases holds: (1) X can be written as a well-ordered union of pre-chains, or (2) X admits a perfect set of pairwise -incomparable elements, and the quotient partial order induced by (X, ) embeds into (2 α, lex ) for some ordinal α, or (3) there is an embedding of 2 ω /E 0 into (X, ) whose range consists of pairwise -incomparable elements. By considering the case where is the diagonal on X, it follows that for any set X exactly one of the following cases holds: (1) X is well-orderable, or (2) X embeds the reals and is linearly orderable, or (3) 2 ω /E 0 embeds into X. In particular, a set is linearly orderable if and only if it embeds into P(α) for some α. Also, ω is the smallest infinite cardinal, and {ω 1, R} is a basis for the uncountable cardinals. Assuming the model has the form L(T, R) for some T ORD, the result is a consequence of ZF + DC R together with the existence of a fine σ-complete measure on P ω1 (R) via an analysis of Vopěnka-like forcing. It is known that in the models not covered by this case, AD R holds. The result then requires more of the theory of determinacy; in particular, that V = OD((< Θ) ω ), and the existence and uniqueness of supercompactness measures on P ω1 (γ) for γ < Θ. As an application, we show that (under the same basic assumptions) Scheepers s countable-finite game over a set S is undetermined whenever S is uncountable. Contents 1. Introduction 2 2. Preliminaries 5 3. AD + 19 4. The dichotomy theorem 21 5. The countable-finite game in natural models of AD + 27 6. Questions 29 References 30 2010 Mathematics Subject Classification. 03E60, 03E25, 03C20. Key words and phrases. Determinacy, AD +, AD R, -Borel sets, ordinal determinacy, Vopěnka forcing, Glimm-Effros dichotomy, countable-finite game. 1 c 2010 American Mathematical Society

2 ANDRÉS EDUARDO CAICEDO AND RICHARD KETCHERSID 1. Introduction This paper deals with consequences of the strengthening AD + of the axiom of determinacy AD for the general theory of sets, not just for sets of reals or sets of sets of reals. Particular versions of our results were known either in L(R) or under the additional assumption of AD R. They can be seen as generalizations of well-known facts in the theory of Borel equivalence relations. We consider natural models of AD +, namely, those that satisfy V = L(P(R)), although our results apply to a slightly larger class of models. The special form of V is used in the argument, not just consequences of determinacy. Although an acquaintance with determinacy is certainly desirable, we strive to be reasonably self-contained and expect the paper to be accessible to readers with a working understanding of forcing, and combinatorial and descriptive set theory. We state explicitly all additional results we require, and provide enough background to motivate our assumptions. Jech [15] and Moschovakis [25] are standard sources for notation and definitions. For basic consequences of determinacy, some of which we will use without comment, see Kanamori [16]. 1.1. Results. Our main result can be seen as a simultaneous generalization of the Harrington- Marker-Shelah [10] theorem on Borel orderings, the Dilworth decomposition theorem of Foreman [7], the Glimm-Effros dichotomy of Harrington-Kechris-Louveau [9], and the dichotomy theorem of Hjorth [13]. Recall that a pre-partial ordering on a set X (also called a quasi-ordering on X) is a binary relation that is reflexive and transitive, though not necessarily anti-symmetric. Recall that E 0 is the equivalence relation on 2 ω defined by xe 0 y n m n ( x(m) = y(m) ). Theorem 1.1. Assume AD + holds and either V = L(T, R) for some T ORD, or else V = L(P(R)). Let (X, ) be a pre-partially ordered set. Then exactly one of the following holds: (1) X is a well-ordered union of -pre-chains. (2) There are perfectly many -incomparable elements of X, and there is an order preserving injection of the quotient partial order induced by X into (2 α, lex ) for some ordinal α. (3) There are 2 ω /E 0 many -incomparable elements of X. The argument can be seen in a natural way as proving two dichotomy theorems, Theorems 1.2 and 1.3. Theorem 1.2. Assume AD + holds and either V = L(T, R) for some T ORD, or else V = L(P(R)). Let (X, ) be a pre-partially ordered set. Then either: (1) There are perfectly many -incomparable elements of X, or else (2) X is a well-ordered union of -pre-chains. Theorem 1.3. Assume AD + holds and either V = L(T, R) for some T ORD, or else V = L(P(R)). Let (X, ) be a partially ordered set. Then either: (1) There are 2 ω /E 0 many -incomparable elements of X, or else (2) There is an order preserving injection of (X, ) into (2 α, lex ) for some ordinal α.

A TRICHOTOMY THEOREM IN NATURAL MODELS OF AD + 3 It is easy to see that R injects into 2 ω /E 0, and it is well-known that, under determinacy, ω 1 does not inject into R, and 2 ω /E 0 is not linearly orderable and therefore cannot embed into any linearly orderable set. This shows that the cases displayed above are mutually exclusive. Theorem 1.2 generalizes a theorem of Foreman [7] where, among other results, it is shown (in ZF + AD + DC R ) that if is a Suslin/co-Suslin pre-partial ordering of R without perfectly many incomparable elements, then R is a union of λ-many Suslin sets, each pre-linearly-ordered by, where λ is least such that both and its complement are λ-suslin. By considering the case = {(x, x) : x X}, the following corollary, a generalization of the theorem of Silver [29] on co-analytic equivalence relations, follows immediately: Theorem 1.4. Assume AD + holds and either V = L(T, R) for some T ORD, or else V = L(P(R)). Let X be a set. Then either: (1) R embeds into X, or else (2) X is well-orderable. The corollary gives us the following basis result for infinite cardinalities: Corollary 1.5. Assume AD + holds and either V = L(T, R) for some T ORD, or else V = L(P(R)). Let S be an infinite set. Then: (1) ω embeds into S. (2) If κ is a well-ordered cardinal, and S is strictly larger than κ, then either κ + or κ R embeds into S. In particular, ω 1 and R form a basis for the uncountable cardinals. Note that there are no assumptions in Theorems 1.2 1.4 on the set X. If, in Theorem 1.4, the set X is a quotient of R by, say, a projective equivalence relation, one can give additional information on the length of the well-ordering. This has been investigated by several authors including Harrington-Sami [11], Ditzen [5], Hjorth [12], and Schlicht [28]. Theorems 1.2 and 1.4 were our original results, and we consider Theorem 1.2 the main theorem of this paper. After writing a first version of the paper, we found Hjorth [13], where the version of Theorem 1.4 for L(R) is attributed to Woodin. Hjorth [13] investigates in L(R) what happens when alternative 1 in Theorem 1.4 holds but the quotient R/E 0 does not embed into X; much remains to be explored in this area. We remark that the argument of Hjorth [13] easily combines with our techniques, so we in fact have Theorem 1.3, a simultaneous generalization of further results in Foreman [7], and the main result in Hjorth [13]. The following corollary is immediate: Corollary 1.6. Assume AD + holds and either V = L(T, R) for some T ORD, or else V = L(P(R)). Let X be a set. Then either: (1) 2 ω /E 0 embeds into X, or else (2) X embeds into P(α) for some ordinal α. In particular: Corollary 1.7. Assume AD + holds and either V = L(T, R) for some T ORD, or else V = L(P(R)). Then a set is linearly orderable if and only if it embeds into P(α) for some ordinal α.

4 ANDRÉS EDUARDO CAICEDO AND RICHARD KETCHERSID Since it is slightly easier to follow, we arrange the exposition around the proof of Theorem 1.4, and then explain the easy adjustments to the argument that allow us to obtain Theorem 1.2, and the modifications required to the argument in Hjorth [13] to obtain Theorem 1.3. Weak versions of some of these results were known previously in the context of AD R. It is thanks to the use of -Borel codes in our arguments that we can extend them in the way presented here. As an application of our results, we show: Theorem 1.8. Assume AD + holds and either V = L(T, R) for some T ORD, or else V = L(P(R)). Then the countable-finite game CF (S) is undetermined for all uncountable sets S. This is a slightly amusing situation in that we have a family of games that are obviously determined under choice, but are undetermined in the natural models of determinacy. Theorem 1.8 seems of independent interest, since it is still open whether, under choice, player II has a winning 2-tactic in CF (R). Theorem 1.8 seems to indicate that the answer to this question only depends on the cardinal c rather than on any particular structural properties of the set of reals. We also present detailed proofs of two additional results, not due to us. First, directly related to our approach is Woodin s theorem characterizing the -Borel sets: Theorem 1.9 (Woodin). Assume ZF + DC R + µ is a fine σ-complete measure on P ω1 (R). Then a set of reals A is -Borel iff A L(S, R), for some S ORD. For models of AD + of the form L(T, R) for some T ORD, Theorems 1.2 and 1.3 are in fact consequences of the assumptions of Theorem 1.9, this we establish via an analysis of -Borel codes by means of Vopěnka-like forcing. In the models not covered by this case, AD R holds, and the results require two additional consequences of determinacy due to Woodin, namely, that V = OD((< Θ) ω ), and the uniqueness of supercompactness measures on P ω1 (γ) for γ < Θ. We omit the proofs of these two facts. Second, we also present a proof of the following result of Jackson: Theorem 1.10 (Jackson). Assume AC ω (R). Then there is a countable pairing function, i.e, a map F : [P(R)] ω P(R) satisfying: (1) F (A) is independent of any particular way A is enumerated, and (2) Each A A is Wadge-reducible to F (A). It is because of Theorem 1.10 that our approach to Theorem 1.2 in the AD R case is different from the approach when V = L(T, R) for some T ORD. 1.2. Organization of the paper. Section 2 provides the required general background to understand our results, and includes a brief (and perhaps overdue) motivation for AD +, a quick discussion of

A TRICHOTOMY THEOREM IN NATURAL MODELS OF AD + 5 the known methods for obtaining natural models of determinacy, and a description of Scheepers s countable-finite game. In Section 3 we state without proofs some specific consequences of AD + that our argument needs. We also prove Jackson s Theorem 1.10. In Section 4 we prove Woodin s Theorem 1.9, and the dichotomy Theorem 1.4. The argument divides in a natural way into two cases, according to whether V = L(T, R) for some T ORD, or V = L(P(R)). In the latter case, we may also assume AD R, that we use to derive the result from the former case. The argument in the AD R case was suggested by Hugh Woodin. We also explain how to modify the argument to derive our main result, Theorem 1.2, and sketch how to extend the argument in Hjorth [13] to prove Theorem 1.3. The deduction of Corollary 1.7 from the argument of Theorem 1.3 is standard. In Section 5 we analyze the countable-finite game CF (S) in ZF, and use the dichotomy Theorem 1.4 to show that in models of AD + of the forms stated above, the game is undetermined for all uncountable sets S. Since trivially player II has a winning strategy if S is countable, this provides us with a complete analysis of the game in natural models of AD +. We have written this section in a way that readers mainly interested in this result, can follow the argument without needing to understand the proofs of our main results. Finally, in Section 6 we close with some open problems. 1.3. Acknowledgments. We want to thank Marion Scheepers, for introducing us to the countable-finite game, which led us to the results in this paper; Steve Jackson, for allowing us to include in Subsection 3.3 his construction of a pairing function; Matthew Foreman, for making us aware of Foreman [7], which led us to improve Theorem 1.4 into Theorem 1.2; and Hugh Woodin, for developing the beautiful theory of AD +, for his key insight regarding the dichotomy Theorem 1.4 in the AD R case, and for allowing us to include a proof of Theorem 1.9. The first author also wants to thank the National Science Foundation for partial support through grant DMS-0801189. 2. Preliminaries The purpose of this section is to provide preliminary definitions and background. In particular, we present a brief discussion of AD + in Subsection 2.2, of two methods for obtaining models of determinacy in Subsection 2.3, and of the countable-finite game in Subsection 2.5. 2.1. Basic notation. ORD denotes the class of ordinals. Whenever we write S ORD, it is understood that S is a set. Given a set X, we endow X ω with the (Tychonoff s) product topology of ω copies of the discrete space X, so basic open sets have the form [s] = {f X ω : s f}, where s X <ω. This will always be the case, even if X is an ordinal or carries some other natural topology. R will always mean Baire space, ω ω, that is homeomorphic to the set of irrational numbers.

6 ANDRÉS EDUARDO CAICEDO AND RICHARD KETCHERSID Definition 2.1. A tree T on a finite product i<n X i (typically for us, n = 1 or 2) is a subset of ( i<n X i) <ω that is closed under restrictions. It is customary to identify T with a subset of i<n (X<ω i ) such that: (1) Whenever (p i : i < n) T, then all p i, i < n, have the same length. (2) T is closed under restrictions, in the following sense: If p = (p i : i < n) T and m is smaller than the common length of the p i, i < n, then p m := (p i m : i < n) T. If T is a tree on X Y and x X <ω, then T x = {y Y <ω : (x, y) T } and if x X ω, then T x = T x n, n so T x is a tree on Y. We denote by [T ] the set of infinite branches through T and, if T is a tree on X Y, then p[t ] = {f X ω : g Y ω ( (f, g) [T ] ) } = {f : T f is ill-founded}. As usual, an infinite branch through T is a function f : ω T such that for all n, f n T. 2.1.1. Games. We deal with infinite games, all following a similar format: For some (fixed) set X, two players I and II alternate making moves for ω many innings, with I moving first. In each move, the corresponding player plays an element of X: I x 0 x 2... II x 1 x 3 (Specific games may impose restrictions on what elements are allowed as the play progresses.) This way both players collaborate to produce an element x = x 0, x 1, x 2,... of X ω. Given A X ω, we define the game X (A) by following the format just described, and declaring that player I wins iff x A. A strategy is a function σ : X <ω X. Player I follows the strategy σ iff each move of I is dictated by σ and the previous moves of player II: I σ( ) σ( x 0 ) σ( x 0, x 1 ) II x 0 x 1... Similarly one defines when II follows σ. A strategy σ is winning for I in a game on X iff, for all x = x 0, x 1,... X ω, player I wins the run σ x of the game, produced by I following σ against player II, who plays x bit by bit. Similarly we define when σ is winning for II. We say that a game is determined when there is a winning strategy for one of the players. When the game is X (A) for some A X ω, it is customary to say that A is determined. Definition 2.2 (AD). In ZF, the axiom of determinacy, AD, is the statement that all A R are determined.

A TRICHOTOMY THEOREM IN NATURAL MODELS OF AD + 7 A standard consequence of AD is the perfect set property for sets of reals: Any A R is either countable or contains a perfect subset. It follows that AD is incompatible with the existence of a well-ordering of the reals and, in fact, with the weaker statement ω 1 R, that ω 1 injects (or embeds) into R. Since determinacy contradicts the axiom of choice, it should be understood as holding not in the universe V of all sets but rather in particular inner models, such as L(R). When our results below assume, for example, that V = L(P(R)) and that AD holds, this could then be understood as a result about all inner models M of ZF that satisfy AD + V = L(P(R)). 2.2. AD +. At first the study of models of determinacy might appear to be a strange enterprise. However, as the theory develops, it becomes clear that one is really studying the properties of definable sets of reals. The notion of definability is inherently vague; however, under appropriate large cardinal assumptions, any reasonable notion of A is a definable set of reals is equivalent to A is in an inner model of determinacy containing all the reals. Thus the study of properties of definable sets of reals becomes the focus. 2.2.1. The theory AD +. AD + is a strengthening of AD. The theory of models of AD + is due to Woodin, see for example Woodin [34, Section 9.1]. All unattributed results and definitions in this section are either folklore, or can be safely attributed to Woodin. The starting point for this study is the collection of Suslin sets. Definition 2.3. A set A X ω is κ-suslin iff A = p[s] for some tree S on X κ. A set A is co-κ-suslin if X ω \ A is κ-suslin and we say that A is Suslin/co- Suslin if A is both κ-suslin and co-κ-suslin for some κ. That A is κ-suslin is also expressed by saying that A has a κ-(semi)-scale. In this paper, we have no use for scales other than the incumbent Suslin representation, so we say no more about them. Let S λ = {A R : A is λ-suslin}. Being Suslin is obviously one notion of being definable, and the classically studied definable sets of reals are all Suslin assuming enough determinacy or large cardinals. Actually, choice implies that all sets of reals are Suslin, so under choice one actually studies which sets of reals are in S λ for specific cardinals λ. Without choice, it is not necessarily the case that all sets of reals are Suslin. Definition 2.4. κ is a Suslin cardinal iff S κ \ λ<κ S λ. For example, one can prove in ZF that the first two Suslin cardinals are ω and ω 1. Also, S ω = Σ 1 1, the class of projections of closed subsets of R 2 ; note that the notion of Σ 1 1 sets also makes sense for subsets of R n for n > 1. Assuming some 1 determinacy, then S ω1 = Σ 1 2, the class of projections of complements of 1 sets. It is a classical theorem of Suslin that A is Borel is equivalent Σ to A is ω-suslin/co-suslin. Being Borel is a notion of definability which is obviously extendible by taking longer well-ordered unions. This leads to the notion of -Borel sets, that we describe carefully below, in 2.2.3.

8 ANDRÉS EDUARDO CAICEDO AND RICHARD KETCHERSID For now, define A is -Borel with code (φ, S) to mean that S ORD, φ is a formula in the language of set theory and, for any x R, x A L[S, x] = φ(s, x). Clearly, if T witnesses that A is Suslin, then T also witnesses that A is -Borel, since x A L[T, x] = T x is ill-founded. There are multiple senses in which a code for A is easy to calculate from A, assuming that A is -Borel. One of these will be discussed later, see Theorem 3.4 and 4.1.1, and another is given by Theorem 2.5 below. First, we need a couple of basic notions. Define Θ = sup{ : is a pre-well-ordering of a subset of R}, where is the rank of the pre-well-ordering. Equivalently, Θ = sup{α : f : R onto Suppose that A R n for some n ω, and define Σ 1 1(A) as the smallest collection of subsets of R m with m varying in ω, that contains A and is closed under integer quantification, finite unions and intersections, continuous reduction, and existential real quantification. As usual, define 1(A) to be the class of complements 1 Π 1 of Σ 1 1(A) sets, 2(A) = Σ R 1 1(A), etc. Each of these classes has a canonical universal set U Π n(a). 1 See Moschovakis [25] for notation, the definition of universality, and this fact. If is a pre-well-order of length γ, then we say that S γ is Σ 1 n( ) in the codes iff there is a real x such that for ξ γ, α}. ξ S y [ y = ξ and U 1 n( )(x, y) ]. The Moschovakis Coding Lemma, see Moschovakis [25], states that, under determinacy, given any pre-well-order of R of length γ, any S γ is Σ 1 1( ) in the codes. This yields that if M and N are transitive models of AD with the same reals, and γ < min{θ M, Θ N }, then P(γ) M = P(γ) N. We then have the following regarding -Borel codes. Theorem 2.5 (Woodin). Assume AD and that A is -Borel. Then there is a 1 γ < Θ, a pre-well-order in 2(A) of length γ, and a code S γ for A. By the coding Lemma, S is Σ 1 Π 1 1( ) in the codes. So S is 3(A) in the codes. Σ In particular, if M and N are transitive models of AD with R M = R N, and A M is Suslin (or just -Borel) in N, then A is -Borel in M, although it need not be the case that A is also Suslin in M. The following is essentially contained in results of Kechris-Kleinberg-Moschovakis-Woodin [17], see also Jackson [14]. Theorem 2.6. Assume AD, and suppose that λ < Θ and that A λ ω is Suslin/co-Suslin. Then the game λ (A) is determined. Suppose that M is a transitive model of AD, λ < Θ M, and f : λ ω R is in M and continuous. Let A be a set of reals in M, and consider the (A, f)-induced game on λ, λ (f 1 [A]). Suppose moreover that there is a transitive model N of AD with the same reals as M, and such that A is Suslin/co-Suslin in N. Then, by Theorem

A TRICHOTOMY THEOREM IN NATURAL MODELS OF AD + 9 2.6, in N, λ (f 1 [A]) is determined and hence, by the Coding Lemma, this game is determined in M, since the winning strategy can be viewed as a subset of λ. Finally, recall that Suslin subsets R R 2 can be uniformized, see Moschovakis [25], so that there is a partial function f : R R such that whenever x R and there is a y R with xry then, in fact, x dom(f) and xrf(x). Suppose that M is a transitive model of AD, and that R R 2 is a relation in M such that for any x R there is a y R such that xry. If there is a transitive model N of AD, with the same reals as M, and such that R is Suslin in N, then R is uniformizable in N. If f is a uniformizing function for R in N, then for any real x 0 N there is then a real x N coding the sequence x n : n < ω where x n+1 = f(x n ) for all n ω. Since M and N have the same reals, then x and therefore x n : n < ω are in M. This shows that DC R holds in M (see 2.2.2 below for the definition of DC R ). In summary, we have that if M is a transitive model of AD such that for each A P(R) M, there is a transitive N such that: (1) N models AD, (2) N has the same reals as M and, (3) in N, A is Suslin, then the following hold in M: DC R. All sets of reals are -Borel. For all ordinals λ < Θ M, all continuous functions f : λ ω R, and all A R, the (A, f)-induced game on λ is determined. This situation is axiomatized by AD +. of Definition 2.7 (Woodin). Over the base theory ZF, AD + is the conjunction DC R. All sets of reals are -Borel. < Θ-ordinal determinacy, i.e., all (A, f)-induced games on ordinals λ < Θ are determined, for any A R and any continuous f : λ ω R. The following is a consequence of the preceding discussion. Theorem 2.8. If M is a transitive model of ZF + AD such that every set of reals in M is Suslin in some transitive model N of ZF + AD with the same reals, then M = AD +. In fact, in Theorem 2.8, it suffices that M and N satisfy the restriction of ZF to Σ n sentences, for an appropriate sufficiently large value of n. Remark 2.9. Suppose that M and N are transitive models of AD with the same reals. Let θ = min{θ M, Θ N }. Then, by the Coding Lemma, ( P(γ) ) M ( = P(γ) ) N. γ<θ In particular, if A M N is a set of reals, and A is κ-suslin in N, for some κ < Θ M, then A is κ-suslin in M as well. Recall that Wadge-reducibility of sets of reals is given by γ<θ A W B

10 ANDRÉS EDUARDO CAICEDO AND RICHARD KETCHERSID iff there is a continuous function f : R R such that A = f 1 [B]. It is a basic consequence of determinacy that W is well-founded. We can then assign a rank to each set of reals. The rank of W itself is exactly Θ. Obviously, a continuous reduction can be coded by a real. With M and N as above, we then have that if A M N is a set of reals, then A M W = A N W. It follows that if A is not Suslin in M but it is Suslin in N, then P(R) M P(R) N and Θ M < Θ N. A benefit of considering AD + rather than AD is that much of the fine analysis of L(R) under the assumption of determinacy actually lifts to models of the form L(P(R)) under the assumption of AD +. Whether AD + actually goes beyond AD is a delicate question, still open. We will briefly touch on this below. 2.2.2. DC R. Recall that DC R, or DC ω (R), is the statement that whenever R R 2 is such that for any real x there is a y with xry, then there is a function f : ω R such that for all n, f(n)rf(n + 1). It is easy to see that this is equivalent to the claim that any tree T on R with no end nodes has an infinite branch. Two straightforward (and well-known) observations are worth making: First, in ZF, assume that DC R holds and that T ORD. Then DC R holds in L(T, R). Second, if DC R holds in L(T, R) then, in fact, L(T, R) satisfies the axiom of dependent choices, DC. It is shown in Solovay [30] that for models satisfying V = L(P(R)) and in fact, more generally, for models of V = OD(P(R)), if AD + DC R holds, then cf(θ) > ω = DC. Under AD, there are interesting relationships and variations of DC R, due to the existence of certain measures. Let D denote the set of Turing degrees. A set A D is a cone iff there is an a D such that A = {b D : a T b}, where T denotes the relation of Turing reducibility. Define the Martin measure µ M on D, by A µ M A contains a Turing cone. Martin proved that µ M is a σ-complete measure on D. We have: DC = ORD/µ M is well-founded = ω1 /µ M is well-founded = DC R. The first and second implications are trivial. Here is a quick sketch of the third: Lemma 2.10 (Woodin). Over ZF, assume that µ M is a measure, and that ω1 /µ M is well-founded. Then DC R holds. Proof. Let T be a tree on R. For d D, let T d be the tree T restricted to nodes recursive in d. T d is in essence a tree on ω and, since DC ω (ω) certainly holds, T d is ill-founded iff T d has an infinite branch. If T d is ill-founded for any d, then there is an infinite branch through T, so assume that all trees T d are well-founded. For each x R <ω, we can define a partial function by h x : D ω 1 h x (d) = rk Td ( x),

A TRICHOTOMY THEOREM IN NATURAL MODELS OF AD + 11 leaving h x (d) undefined if x / T d. Note that h x (d) is defined for µ M -a.e. degree d. By assumption, [h x ] µm is an ordinal α x, and the map x α x ranks the original tree T and hence T is not a counterexample to DC R. Clearly, in this argument, the Turing degree measure could be replaced by any σ-complete, fine measure µ on P ω1 (R) satisfying that ω 1 /µ is well-founded. Under AD + DC R we actually have the equivalence ORD/µM is well-founded DC R. The left-hand side of this equivalence was part of Woodin s original formalization of AD +. There are models of AD + + cf(θ) = ω. In these models, DC fails, so just the well-foundedness of ultrapowers by fine measures on P ω1 (R) does not give DC. 2.2.3. -Borel sets. Essentially the -Borel sets are the result of extending the usual Borel hierarchy by allowing arbitrary well-ordered unions. Work in ZF. Without choice it is better to work with codes for sets (descriptions of their transfinite Borel construction) rather than with the sets themselves (the output of such a construction), hence an -Borel set is any set with an - Borel code. For example, it might be the case that for all α < γ, A α is -Borel, but there is no sequence of codes c α and hence α<γ A α might not be -Borel. There are several equivalent definitions of -Borel codes. For definiteness, we present an official version, and then some variants, and leave it up to the reader to check that the notions are equivalent, and even locally equivalent when required. Definition 2.11. Fix a countable set of objects N = {, } {ṅ : n ω} with N disjoint from ORD; e.g., = (0, 0), = (0, 1), and ṅ = (1, n) would suffice. The -Borel codes (BC) are defined recursively by: T BC iff one of the following holds: T = ṅ. T = α<κ T α = {, α s : α < κ and s T α } where each T α BC. T = S = { s : s S} where S BC. Hence a code is essentially a well-founded tree on ORD N, and we will identify -Borel codes with these trees without comment. Set BC κ = BC {T : T is a well-founded tree of rank < κ}. For κ a limit ordinal, BC κ is closed under finite joins. If cf(κ) > ω, then BC κ is σ-closed and, if κ is regular, then BC κ is < κ-closed. Clearly for regular κ, BC κ = BC H(κ). Definition 2.12. A set of reals is -Borel iff it is the interpretation of some T BC. We denote this interpretation by A T, and define it by recursion as follows: Aṅ = {x R : x(n 0 ) = n 1 }, where n (n 0, n 1 ) is a recursive bijection between R and R 2. A α<κ Tα = α<κ A T α.

12 ANDRÉS EDUARDO CAICEDO AND RICHARD KETCHERSID A T = R \ A T. The predicates T BC and x A T are Σ 1 and absolute for any model of KP + Σ 1 -separation. (Just KP is not enough, since the code must be well-founded.) Let B denote the collection of -Borel sets, and let B κ be the subset of B consisting of those sets with codes in BC κ. In particular, if ω 1 is regular, then B ω1 is just the algebra of Borel sets. The following gives a few alternate definitions for the -Borel sets. The equivalence of the first three is local in the sense that it is absolute to models of KP+Σ 1 - separation. The equivalence with the fourth one is still reasonably local, certainly absolute to models of ZF, and the definition itself can be formalized in any theory strong enough to allow the definability of the satisfiability relation for the classes L[S, x]. A is -Borel. There is a tree T on κ ω such that A(x) iff player I has a winning strategy in the game T,x given by: Players I and II take turns playing ordinals α i < κ so in the end they play out f κ ω. Player I wins iff (f, x) [T ]. Note that the game T,x is closed for I and hence determined. (In this case T is taken as the code and A T = {x : I has a winning strategy in T,x }.) There is a Σ 1 formula φ (in the language of set theory, with two free variables) and S γ for some γ, such that A(x) L[S, x] = φ(s, x). (Here (φ, S) is taken to be the code and A = A φ,s is the set coded.) There is a formula φ and S γ for some γ, such that A(x) L[S, x] = φ(s, x). (Once again, (φ, S) is taken to be the code and A = A φ,s is the set coded.) It is thus natural to identify codes with sets of ordinals, and we will often do so. For example, as mentioned above, Suslin sets are -Borel. On the other hand, Suslin subsets of R R can be uniformized, while in general there can be nonuniformizable sets in a model of AD +, so it is not true that all -Borel sets are Suslin. Under fairly mild assumptions, being -Borel already entails many of the nice regularity properties shared by the Borel sets. In particular, suppose that S is a code witnessing that A S is -Borel, and suppose that P(P c ) L[S] V = ω, where P c = Add(ω, 1) is the Cohen poset (essentially ω <ω ). Then A S has the property of Baire. Similarly, if P(P L ) L[S] V = ω, where P L = Ran ω is random forcing, then A S is Lebesgue measurable. In general, if ω1 V is inaccessible in L[S], then A S has all the usual regularity properties. Note that Theorem 1.9 provides us, over the base theory ZF + DC R + there is a fine measure on P ω1 (R), with yet another equivalence for the notion of -Borel; however, we know of no reasonable sense in which this version would be local as the previous ones.

A TRICHOTOMY THEOREM IN NATURAL MODELS OF AD + 13 2.2.4. Ordinal determinacy. AD states that all games on ω are determined. One may wonder whether it is consistent with ZF that, more generally, all games on ordinals are determined. This is not the case; in fact, it is well-known that there is an undetermined game on ω 1. To see this, consider two cases. If AD fails, we are done, and there is in fact an undetermined game on ω. If AD holds, then ω 1 R. Consider the game where player I begins by playing some α < ω 1, and player II plays bit by bit a real coding ω + α. Since any countable ordinal can be coded by a real, it is clear that player I cannot have a winning strategy. Were this game determined, player II would have a winning strategy σ. But it is straightforward to define from σ an uncountable injective sequence of reals, and we reach a contradiction. It follows that some care is needed in the way the payoff of ordinal games is chosen if we want them to be determined, and this is why < Θ-determinacy is stated as above. Note that ordinal determinacy indeed implies determinacy, so AD + strengthens AD. One consequence of ordinal determinacy that we will use is the following: Theorem 2.13 (Woodin). Assume AD +. Then, for every Suslin cardinal κ, there is a unique normal fine measure µ κ on P ω1 (κ). In particular, µ κ OD. If κ is below the supremum of the Suslin cardinals, this follows from Woodin [33], where games on ordinals are simulated by real games, in particular, giving the result under AD R (which is the case that interests us). For the AD + result, Woodin s argument must be integrated with the generic coding techniques in Kechris-Woodin [19] to produce ordinal games that are determined under AD +. The result is that the supercompactness measure coincides with the weak club filter, where S P ω1 (κ) a is weak club iff S = κ and, whenever σ 0 σ 1 are in S, then i ω σ i S. Let κ be a Suslin cardinal. For any γ < κ, define µ γ = π κ,γ (µ κ ) where π κ,γ : P ω1 (κ) P ω1 (γ) is defined by σ σ γ. This gives a canonical sequence of ω 1 -supercompactness measures on all γ less than the supremum of the Suslin cardinals. 2.2.5. AD R. Over ZF, AD R is the assertion that for all A R ω, the game R (A) is determined. DC R is an obvious consequence of AD R, and Woodin has shown that AD R yields that all sets of reals are -Borel. However, as far as we know, the only proof of AD R = AD + uses an argument of Becker [2] for getting scales from uniformization, and Becker s proof uses DC. The minimal model of AD R does not satisfy DC, but does satisfy AD + ; this requires a different argument basically analysing the strength of the least place where AD + AD + could hold. Woodin has shown from AD R + AD + that all sets are Suslin, without appeal to Becker s argument. At the moment, the lack of a proof (not assuming DC) that AD R = AD +, and hence that AD R = all sets are Suslin, seems to be a weakness in the theory. To make results easy to state, from here on AD R will mean AD R + AD +. Let κ = sup{κ : κ is a Suslin cardinal}. Assuming AD, κ = Θ all sets of reals are Suslin.

14 ANDRÉS EDUARDO CAICEDO AND RICHARD KETCHERSID Theorem 2.14 (Steel, Woodin). The following hold in ZF: (1) AD + DC R implies that the Suslin cardinals are closed below κ. (2) AD R is equivalent to AD + κ = Θ. (3) AD + is equivalent to AD + DC R together with the Suslin cardinals are closed below Θ. (For a proof of the first and second items, and a sketch of the third, see Ketchersid [20].) Thus, at least in the presence of DC R, if there is a model of AD + AD +, then in this model κ < Θ and κ is not a Suslin cardinal. The main open problem in the theory of AD is whether AD does in fact (over ZF) imply AD +. 2.2.6. L(R). It is not immediate even that L(R) = AD AD +. This is the content of the following results: Theorem 2.15 (Kechris [18]). Assume V = L(R) = ZF+AD. Then DC R (and therefore DC) holds. As mentioned previously, in the context of choice, it is automatic that DC holds in L(R), regardless of whether AD does. Woodin has found a new proof of Kechris s result using his celebrated derived model theorem, stated in Subsection 2.3. The basic fine structure for L(R) yields that, working in L(R), if Γ(x) is the lightface pointclass consisting of all sets of reals Σ 1 -definable from x, then Γ(x) = Σ 2 1(x), the collection of all sets A of reals such that y A B R φ(b, x, y) for some Π 1 2 formula φ. As usual, Π 2 1(x) is the collection of complements of Σ 2 1(x) sets, and 2 1(x) is the collection of sets that are both Σ 2 1(x) and Π 2 1(x). Solovay s basis theorem, see Moschovakis [25], goes further to assert that the witnessing set can in fact be chosen to be 2 1(x), that is, x A B 2 1(x) φ(b, x). In Martin-Steel [23], it is shown that, under AD, Σ L(R) 1 has the scale property. For us, this means that every set in Σ L(R) 1 is Suslin. Combining these two results gives that any Σ L(R) 1 fact about a real x has a Suslin/co-Suslin witness. Let n be as in the paragraph following Theorem 2.8. The theory ZF n resulting from only considering those axioms of ZF that are at most Σ n sentences, is finitely axiomatizable. Suppose L(R) failed to satisfy AD +. Then the following Σ L(R) 1 statement holds: M [ R M and M = ZF n + AD +]. By the basis theorem together with the Martin-Steel result, the witness M can be coded by a Suslin/co-Suslin set. Thus M L(R) are two transitive models of ZF n +AD with the same reals, and one can check that each set of reals in M is Suslin in L(R). It follows from Theorem 2.8 that M = AD + and this is a contradiction. This proves: Corollary 2.16. L(R) = AD AD +. Two results that hold for L(R) whose appropriate generalizations are relevant to our results are the fact that in L(R) every set is ordinal definable from a real, and the following:

A TRICHOTOMY THEOREM IN NATURAL MODELS OF AD + 15 Theorem 2.17 (Woodin). L(R) = S Θ (HOD = L[S]). The set S as in Theorem 2.17 is obtained by a version of Vopěnka forcing due to Woodin that can add R to HOD L(R). Variants of this forcing are very useful at different points during the development of the AD + theory, the general version being: Theorem 2.18 (Woodin). Suppose that AD + holds and that V = L(P(R)). Then there is S Θ such that HOD = L[S]. S can be taken to code the Σ 1 -theory of Θ in L(P(R)). If V = L(T, R) for some set T ORD, then S can be obtained by a generalization of the version of Vopěnka forcing hinted at above. The stronger statement that P(R) L(S, R) is false in general. For example, it implies that AD R fails, as claimed in Woodin [34, Theorem 9.22]. 2.3. Obtaining models of AD +. Here we briefly discuss two methods by which (transitive, proper class) models of AD + (that contain al the reals) can be obtained; this illustrates that there is a wide class of natural models to which our results apply: 2.3.1. The derived model theorem. The best understood models of AD + come from a construction due to Woodin, the derived model theorem. In a precise sense, this is our only source of natural models of AD +. The derived model theorem carries two parts, first obtaining models of determinacy from Woodin cardinals, and second recovering models of choice with Woodin cardinals from models of determinacy. Although the full result remains unpublished, proofs of a weaker version can be found in Steel [31, 32] and Koellner- Woodin [21]. Theorem 2.19 (Woodin). (ZFC) Suppose δ is a limit of Woodin cardinals. Let V (R ) be a symmetric extension of V for Coll(ω, < δ), so R = α<δ R V [G α] for some Coll(ω, < δ)-generic G over V. Then: (1) R = R V (R ), V (R ) = AC, and V (R ) = DC iff δ is regular. (2) Define Γ = { A R : A V (R ) and L(A, R ) = AD + }. Then L(Γ, R ) = AD +. Notice that L(Γ, R ) = V = L(P(R)) and that, in particular, the theorem implies Γ. Remark 2.20. If δ as above is singular, then R R V [G]. It is the fact that the theorem admits a converse that makes it the optimal result of its kind, in the sense that it captures all the L(P(R))-models of AD + : Theorem 2.21 (Woodin). Suppose V = L(P(R)) + AD +. There exists P such that, if G is P-generic over V then, in V [G], one can define an inner model N = ZFC such that:

16 ANDRÉS EDUARDO CAICEDO AND RICHARD KETCHERSID (1) ω V 1 is limit of Woodin cardinals in N. (2) N(R V ) is a symmetric extension of N for Coll(ω, < ω V 1 ). (3) V = N(R V ). Remark 2.22. N is not an inner model of V. If it were, every real of V would be in a set-generic extension of a (fixed) inner model of V by a forcing of size < ω V 1. AD prevents this from happening, as it is a standard consequence of determinacy that any subset of ω 1 is constructible from a real. The point here is that to be a symmetric extension is first order, as the following well-known result of Woodin indicates (see Bagaria-Woodin [1] or Di Prisco- Todorčević [4] for a proof): Lemma 2.23 (Woodin). Suppose N = ZFC, let δ be a strong limit cardinal of N, and let σ R. Then N(σ) is a symmetric extension of N for Coll(ω, < δ) iff (1) Whenever x, y σ, then R N[x, y] σ, (2) Whenever x σ, then x is P-generic over N for some P N such that P N < δ, and (3) sup x σ ω N[x] 1 = δ. Let us again emphasize that all the models obtained using the construction described in the derived model theorem satisfy V = L(P(R)), and they also satisfy AD +. 2.3.2. Homogeneous trees. The second method we want to mention is via homogeneously Suslin representations in the presence of large cardinals. We briefly recall the required definitions. The key notion of homogeneous tree was isolated independently by Kechris and Martin 1 from careful examination of Martin s proof of 1-determinacy from a measurable cardinal. Π Definition 2.24. Let 1 n m < ω. For X a set and A X m, let A n := { u n : u A }. Let κ be a cardinal, and let µ and ν be measures on κ n and κ m, respectively. We say that µ and ν are compatible iff A κ m (A ν A n µ) or, equivalently, iff B µ { u κ m : u n B } ν. Definition 2.25. Let T be a tree on ω κ. We say that µ u : u ω <ω is a homogeneity system for T iff (1) For each u ω <ω, µ u is an ω 1 -complete ultrafilter on T u (i.e., T u µ u ), (2) For each u v ω <ω, µ u and µ v are compatible, and (3) For any x R, if x p[t ] and A i µ x i for all i < ω, then there is f : ω κ such that i (f i A i ). We say that T is a homogeneous tree just in case it admits a homogeneity system, and we say it is κ-homogeneous iff it admits a homogeneity system µu : u ω <ω where each µ u is κ-complete.

A TRICHOTOMY THEOREM IN NATURAL MODELS OF AD + 17 Note that if µ is a homogeneity system for T and x / p[t ] then, setting A i = T x i, there is no f such that i (f i A i ). Thus, item 3 of Definition 2.25 gives a characterization of membership in p[t ]. The key fact relating determinacy and the notion of homogeneous trees is the following: Theorem 2.26 (Martin). If A = p[t ] for some homogeneous tree T, then ω (A) is determined. Definition 2.27. A set A R is homogeneously Suslin iff there is a homogeneous tree T such that A = p[t ]. A is κ-homogeneously Suslin (or κ-homogeneous) iff it is the projection of a κ-homogeneous tree. A is -homogeneous iff it is κ-homogeneous for all κ. 1 For example, 1-sets are homogeneously Suslin: For any measurable κ and any 1 Π 1-set A, there is a κ-homogeneous tree T on ω κ with A = p[t ]. Π All the proofs of determinacy from large cardinals have actually shown that the pointclasses in question are not just determined, but consist of homogeneously Suslin sets. Under large cardinal hypotheses, the -homogeneous sets are closed under nice operations. For example: Theorem 2.28 (Martin-Steel [24]). Let δ be a Woodin cardinal. Suppose that A R 2 is δ + -homogeneous and Then B is κ-homogeneous for all κ < δ. B = R A := {x : y ((x, y) / A)}. This allows us to identify, from enough large cardinals, nice pointclasses Γ P(R) such that L(Γ, R) = AD. In fact, although this is not a straightforward adaptation of the sketch presented for L(R), the arguments establishing that sets in Γ are (sufficiently) homogeneous also allow one to show that L(Γ, R) = AD +. Notice that, once again, L(Γ, R) = V = L(P(R)). A posteriori, it follows that these models arise by applying the derived model theorem to a suitable forcing extension of an inner model of V. 2.4. Canonical models of AD +. AD + is essentially about sets of reals; in particular, if AD + holds, then it holds in L(P(R)). We informally say that models of this form are natural and note that, for investigating global consequences of AD +, these are indeed the natural inner models to concentrate on. There are however, other canonical inner models of AD +, typically of the form L(P(R))[X] for various nice X. Niceness here means that the models satisfy an appropriate version of condensation. For example, L(R)[µ] where µ a fine measure on P ω1 (R) which is moreover normal in the sense of Solovay [30]; or L(R)[E] for E a coherent sequence of extenders. We will not consider these more general structures in this paper.

18 ANDRÉS EDUARDO CAICEDO AND RICHARD KETCHERSID As explained in the previous subsection, the best known methods of producing models of determinacy actually give us models of AD + + V = L(P(R)). Of course, not all known models of AD + have a nice canonical form, but they are typically obtained from these models, for example, by going to a forcing extension, as in Woodin s example in Kechris [18] of a model of AD + + AC ω obtained by forcing over L(R). Woodin has shown that any model of AD + of the form L(P(R)) either satisfies V = L(T, R) for some set T ORD, or else it is a model of AD R ; a precise statement will be given in Theorem 3.1 below. This may help explain the hypothesis in the statement of our results in Subsection 1.1. 2.5. The game CF (S). Scheepers [26] introduced the countable-finite game around 1991. It is a perfect information, ω-length, two-player game relative to a set S. We denote it by CF (S). I O 0 O 1... II T 0 T 1 At move n, player I plays O n, a countable subset of S, and player II responds with T n, a finite subset of S. Player II wins iff n O n n T n. Obviously, under choice, player II has a winning strategy. Scheepers [26, 27] investigates what happens when the notion of strategy is replaced with the more restrictive notion of k-tactic for some k < ω: As opposed to strategies, that receive as input the whole sequence of moves made by the opponent, in a k-tactic, only the previous k moves of the opponent are considered. Tactics being much more restrictive, additional conditions are then imposed on the players: Player I must play increasing sets: O 0 O 1.... Player II wins iff n O n = n T n. This setting is not completely understood yet. In ZFC: Player I does not have a winning strategy, and therefore no winning k- tactic for any k. Player II does not have a winning 1-tactic for any infinite S. (Scheepers [27]) Player II has a winning 2-tactic for S if S < ℵ ω. (Koszmider [22]) Under reasonably mild assumptions (namely, that all singular cardinals κ of cofinality ω are strong limit cardinals and carry a very weak square sequence in the sense of Foreman-Magidor [8]), player II has a winning 2-tactic for any S. (Koszmider [22]) It is still open whether (in ZFC) player II has a winning 2-tactic for CF (ℵ ω ) or for CF (R). In view of the open problems just mentioned, it is natural to consider the countable-finite game in the absence of choice, to help clarify whether AC really plays a role in these problems. This was our original motivation for showing the dichotomy Theorem 1.4, so that we could deduce Theorem 1.8 explaining that in natural models of determinacy, the game CF (S) is undetermined for all uncountable sets S.

A TRICHOTOMY THEOREM IN NATURAL MODELS OF AD + 19 3. AD + We work in ZF for the remainder of the paper. In this section we state without proof some consequences of AD + that we require. 3.1. Natural models of AD +. To help explain the hypothesis of Theorems 1.1 1.3, we recall the following result. Given a set T ORD, the T -degree measure µ T is defined as follows: First, say that a µt b for a, b R iff a L[T, b], and let the µ T -degree of a be the set of all b such that a µt b µt a. Letting D µt denote the set of µ T degrees, we define cones and the measure µ T just as they where defined for the set D of Turing degrees in 2.2.2. The same proof showing that, under determinacy, µ M is a measure on D gives us that µ T is a measure on D µt for all T ORD. Theorem 3.1 (Woodin). Assume AD + +V = L(P(R)) and suppose that κ < Θ. Let T (ω κ ) <ω be a tree witnessing that κ is Suslin. Then V = L(T, R) where T = x T/µ T. This immediately gives us, via Theorem 2.14: Corollary 3.2 (Woodin). Assume AD + + V = L(P(R)). Then either V is a model of AD R, or else V = L(T, R) for some T ORD. On the other hand, no model of the form L(T, R) for T ORD can be a model of AD R. Ultrapowers by large degree notions, as in the theorem above, will be essential towards establishing our result in the L(T, R) case. For models of AD R, a different argument is required, and the following result is essential to our approach: Theorem 3.3 (Woodin). Assume AD R + V = L(P(R)). Then where (< Θ) ω = γ<θ γω. V = OD((< Θ) ω ), 3.2. Closeness of codes to sets. There are a couple of ways in which -Borel codes are close to the sets they code. One way is expressed by Theorem 2.5 above. More relevant to us is the following: Theorem 3.4 (Woodin). Assume AD + + V = L(P(R)). Let T ORD and let A R be OD T. Then A has an OD T -Borel code. Just as an example of how determinacy can be separated from its structural consequences, the preceding theorem essentially is proved by showing: Theorem 3.5 (Woodin). Suppose that V = L(P(γ)) = ZF + DC and that µ is a fine measure on P ω1 (P(γ)) in V. Then, for all T ORD and A R, if A OD T,µ, then A is -Borel and has an -Borel code in HOD T,µ. Fact 3.6. Under AD, there is an OD measure on P ω1 (P(γ)) for all γ < Θ. As a corollary, if AD + holds, V = L(P(γ)) for γ < Θ, and A OD T P(R), then A has an OD T -Borel code. This almost gives Theorem 3.4 since, assuming AD + + V = L(P(R)), we have V = L( γ<θ P(γ)). On the other hand, note that Theorem 3.4 is not immediate from Theorem 1.9, even if V = L(S, R).

20 ANDRÉS EDUARDO CAICEDO AND RICHARD KETCHERSID 3.3. A countable pairing function on the Wadge degrees. Our original approach to the dichotomy Theorem 1.2 required the additional assumption that cf(θ) > ω. Both when trying to generalize this approach to the case cf(θ) = ω, and while establishing Theorem 1.8 on the countable-finite game in general, an issue we had to face was whether countable choice for finite sets of reals could fail in a model of AD R. That this is not the case follows from the existence of a pairing function. Steve Jackson found (in ZF) an example of such a function. Although this is no longer relevant to our argument, we believe the result is interesting in its own right. Below is Jackson s construction. Theorem 3.7 (Jackson). In ZF, there is a function F : P(R) P(R) P(R) satisfying: (1) F (A, B) = F (B, A) for all pairs (A, B), and (2) Both A and B Wadge-reduce to F (A, B). Proof. If A = B, simply set F (A, B) = A. If A B or B A, set F (A, B) = (0 S) (1 T ) where S is the smaller of A, B, and T is the larger. Here, 0 S = {0 a : a S} and similarly for 1 T. If A \ B and B \ A are both non-empty, we proceed as follows: Let X(A, B) R Z be defined by saying that, if f : Z R, then f X(A, B) iff there is an i such that f(i) A \ B (or B \ A) and, for each j, f(j) A if j i is even, and f(j) B if j i is odd (and reverse the roles of A, B here if f(i) B \ A). The set X(A, B) is an invariant set (with respect to the shift action of Z on R Z ), and X(A, B) = X(B, A). (Thus the points of A \ B and B \ A have to occur at places of different parity; while points of A B can occur anywhere.) Given X(A, B), we can compute A (and also B) as follows: Fix z A \ B. Then x A iff f X(A, B) i j (f(i) = z and f(j) = x and j i is even). This shows that A is Σ 1 1(X(A, B)). If we replace X(A, B) with X (A, B), the Σ 1 1- jump of X(A, B), then A is Wadge reducible to X(A, B). Finally, we use that there is a Borel bijection between R Z and R, and define F (A, B) as the image of X (A, B) under this map. As pointed out by Jackson, essentially the same argument shows the following; recall that AC ω (R) is a straightforward consequence of determinacy, so Theorem 3.8 applies in models of AD: Theorem 3.8 (Jackson). Assuming ZF + AC ω (R), there is a countable pairing function. Proof. Let A i : i ω be a sequence of distinct sets of reals. Call f (R 2) ω n-honest iff, whenever f(i) = (x, k), then k = 1 x A n, so f is n-honest iff it is a countable approximation to the characteristic function of A n. Let B = {f : n (f is n-honest)}.